Te Kete o Karaitiana Taiuru (Blog)

The current coalition government has announced the elimination of approximately 8,700 public sector positions or 14% of the public workforce alongside NZ$2.4 billion in spending reductions over three years, with AI deployed as the replacement mechanism. The PSA has been direct in its assessment: AI is being used as a justification for politically driven job cuts, and no government anywhere in the world has replaced workers at this scale with AI.

The PSA warns that experienced staff and the institutional knowledge they carry including years of relationship-building with communities risk being permanently lost, with Māori, rural communities and vulnerable whānau disproportionately affected as services become harder to navigate.

But the full picture is considerably darker than the job numbers suggest. The deployment of AI into government services affecting Māori sits atop three deeper, compounding crises: Māori invisibility in AI systems, structural and historical bias baked into training data, and the large-scale appropriation of mātauranga Māori and te reo without consent, compensation, or governance.

 

Māori Invisibility

The first and most foundational problem is that Māori do not exist, or exist only partially and distortedly, inside most AI systems deployed by government.

AI and machine learning algorithms depend entirely on data for their development. When training data is incomplete, unrepresentative, or culturally biased, AI systems perpetuate harmful stereotypes and make inaccurate predictions. Māori communities face risks because their data is often underrepresented or misinterpreted in mainstream datasets.

Invisibility in training data produces two distinct but related harms. The first is direct misrecognition: Māori worldviews, values, and knowledge systems are distinct and cannot be adequately understood through Western data models. AI systems trained without Māori cultural input may fail to recognise the significance of Māori practices, beliefs, and customs in healthcare, for instance, AI may not address the unique needs of Māori patients or incorporate traditional healing practices, potentially leading to sub-optimal care or harm.

The second harm is the silent imposition of a default: when Māori are absent from training data, the system does not produce a neutral output, it produces a Pākehā-normative output. Every prediction, recommendation, or automated decision that treats the dominant cultural model as the universal one is an act of structural discrimination, even when no discriminatory intent exists at the point of design.

Health New Zealand’s own guidance acknowledges this plainly: generative AI is most likely to have been trained on data that under-represents or misrepresents minority populations including Māori, and AI taught on these biased data sources may reinforce social and health inequities. This acknowledgement sits in guidance material; it has not produced binding requirements on procurement decisions or deployment standards.

The government’s preferred AI providers are predominantly US-based companies subject to US law, not New Zealand companies subject to New Zealand law. With New Zealand’s light-handed regulatory approach, there is no legal requirement that these systems engage with te reo Māori, tikanga, or Māori data sovereignty at any point in their design or deployment.

The invisibility of Māori in AI systems, therefore, is not an accident of history that better datasets will eventually correct. It is being actively reproduced right now by procurement decisions that prioritise cost and speed over equity and Treaty compliance.

 

Algorithmic Bias: Colonialism Automated at Scale

The data that does exist about Māori in public systems overwhelmingly reflects colonial history of over-policing, welfare system contact, criminal justice involvement, socioeconomic disadvantage and AI systems trained on this data will reproduce and amplify those patterns in automated form.

An AI system used to assess bail risk will be more likely to recommend detention for Māori and Pasifika people even when they pose no greater risk, because historical training data reflects decades of over-policing and discriminatory charging practices. Predictive policing systems already operate on this logic globally, and New Zealand’s justice system is being asked to adopt similar tools.

Scholars in algorithmic fairness have argued that framing these outcomes as “algorithmic bias” understates the problem: what is really at stake are power asymmetries and structural injustices that permeate AI technologies, where unfair outcomes from deployed systems can amount to the encoding and amplification of racism with real and immediate impacts on human beings. “Bias” is a technical word that makes a political problem sound solvable with better data science. For Māori, this framing is dangerous because it implies the harm is incidental rather than structural.

The compounding effects reach across every public service domain. In education, predictive models built on deficit-framed data about Māori learners will identify Māori students as lower-risk for engagement or lower-probability for success reinforcing a self-fulfilling prophecy of lower resource allocation. In social services, algorithmic triage systems trained on historical intervention data will over-flag Māori whānau for scrutiny, replicating the over-surveillance that community advocates have fought for generations. In healthcare, recommendation systems that do not account for the social determinants of health as experienced by Māori will produce clinically “optimal” pathways that are culturally inappropriate and practically inaccessible.

Māori data sovereignty represents a Te Tiriti framework obligation requiring that Māori community’s control, govern, and benefit from data about themselves  and it is this obligation that is bypassed when government agencies procure US-based AI tools without Māori governance requirements.

Appropriation of Mātauranga

Large language models and other AI systems are trained on vast scrapes of internet content. That content includes te reo Māori, Māori oral history transcriptions, iwi websites, cultural narratives, whakapapa records, and mātauranga Māori documented across decades of digitisation work. None of this data was contributed with consent for commercial AI training. None of it is governed by Māori. And the companies that hold it including OpenAI, Google, Microsoft and Anthropic are not subject to New Zealand law or Te Tiriti obligations.

Growing concerns about cultural misappropriation centre on the potential for AI systems to extract, reinterpret, or reproduce mātauranga Māori without proper context or authority. Scholars warn that without appropriate safeguards, AI could contribute to the dilution or misuse of cultural knowledge that has been carefully maintained through whakapapa, tikanga, and community stewardship.

The extraction of mātauranga Māori to train commercial AI models that are then sold back to government agencies at profit is a digital iteration of the same logic that converted Māori land into capital: the removal of something held in common by a people, its transformation into a commodity, and its return in a form that generates profit for others while excluding the original holders from governance or benefit.

AI models trained on Māori data may perpetuate stereotypes, misrepresent Māori culture, or appropriate traditional knowledge without permission is a violation of Indigenous rights and self-determination. There are concerns about the potential loss or commodification of traditional knowledge through its incorporation into AI systems, and the opaque nature of AI systems makes it difficult for Māori to understand how their data is being used or hold developers accountable.

More research is needed to explore the potential epistemic violence that can occur in presuming that AI can create across knowledge systems. What is lost through AI processes, particularly when it involves sacred knowledge which certain Indigenous nations rightly control access to requires serious attention.

Mātauranga Māori is not a collection of facts that can be extracted and recombined. It is a system of knowledge embedded in relationships, whakapapa, wāhi, tikanga, and community. When an AI model ingests karakia, whakataukī, and oral tradition and deploys them without this relational context, it does not transmit Māori knowledge. It destroys it, replacing the living system with a simulacrum that the model’s users may mistake for the authentic thing, another form of digital colonisation. The downstream harm is not only to Māori, it is to anyone who subsequently makes decisions informed by an AI’s distorted representation of what Māori knowledge actually is.

 

Data is a Taonga

The Waitangi Tribunal has established that Māori data is, or has the potential to be, a taonga. Under Article 2 of Te Tiriti, which guarantees rangatiratanga over taonga, Māori have inarguable rights to and over any data about Māori, including data used to train AI. The Tribunal is explicit that failure to seek advice from Māori about Māori digital interests constitutes a breach of Te Tiriti.

For Māori, data is more than a commodity,  it is a taonga, an expression of identity, culture, and whakapapa. The governance of Māori data is directly tied to the protection of cultural security, the preservation of language, and the survival of Māori worldviews. The future of AI must be inclusive, participatory, and Treaty-based, with Māori data sovereignty at its core.

The Waitangi Tribunal’s WAI 2522 findings on the CPTPP were particularly significant: Māori claimants argued that key e-commerce provisions posed significant risks to Māori rights, especially over data sovereignty and the protection of mātauranga Māori. The Tribunal concluded a breach of partnership and active protection principles; no formal recommendations made.

The Waitangi Tribunal’s classification of Māori data as taonga establishes a legal foundation for Māori data sovereignty claims and positions data governance as a Treaty issue requiring Crown consultation and Māori consent. A government that deploys AI systems affecting Māori without satisfying this obligation is in Treaty breach, regardless of its efficiency rationale.

 

Structural Argument

The government’s current programme accelerates this dynamic in three ways simultaneously. It removes the human kaimahi who mediate between algorithmic systems and Māori communities. It funds AI deployment through providers with no Te Tiriti obligations. And it does so while potentially getting ready to dismantle the agencies that were built to hold the Crown to account on Māori outcomes.

The honest framing is that the government is choosing to trade Māori rights, safety, and cultural security for a reduction in the public wage bill and using AI as the vehicle through which that trade is made invisible.

 

Potential Solutions

Solutions to these interlocking problems must operate at multiple levels simultaneously: immediate policy, structural governance, legal architecture, and Māori-led infrastructure investment. They must also be grounded in the principle that Māori data sovereignty is not a consultation requirement to be satisfied on the way to a decision that has already been made, it is a precondition for any legitimate AI deployment affecting Māori.

Legislative Reform

The most urgent gap is the absence of legally enforceable protection. The government’s Public Service AI Framework, released in January 2025, is explicitly non-binding, it sets out expectations but imposes no legal obligations on agencies. This is inadequate to the scale of the problem as I have previously written about.

New Zealand needs binding AI regulation that anchors procurement standards in Treaty obligations, requiring that any AI system deployed in public services affecting Māori must demonstrate Māori data sovereignty compliance, bias auditing, and cultural safety assessments before deployment, not after. Australia has begun moving toward stronger sovereign capability and AI assurance requirements in government procurement, including increased scrutiny of high-risk automated systems and algorithmic accountability.

In my State of the Nation report is explicit on the legislative agenda: align Māori data sovereignty with global Indigenous frameworks including the CARE Principles and OCAP; create independent Māori-led monitoring bodies to audit government and corporate use of Māori data; and establish penalties and remedies for misuse, misrepresentation, or non-compliance with Māori data sovereignty principles.

Treaty compliant Procurement Requirements

Every government AI procurement decision affecting Māori communities should be subject to a formal Māori data sovereignty impact assessment as a condition of contracting, not as optional guidance. Canada’s health AI discussions have articulated the principle clearly: every contract involving AI should include clauses on data residency, intellectual property, auditability, and sovereignty and governments can set the tone by simply refusing to purchase tools that do not meet these standards.

Practically, this means procurement contracts must require data residency within New Zealand jurisdiction; prohibition on reuse of Māori data to train third-party commercial models; mandatory bias auditing against Māori outcomes before and after deployment; and contractual rights for Māori governance bodies to access and review algorithmic decision-making. Meaningful consultation must begin before design and procurement decisions are locked in, agencies must document how Māori input shaped outcomes, and there must be a pathway for Māori to help govern how systems are designed, evaluated, and used, not just commented on after the fact.

Strengthening and Mandating the Algorithm Charter

An overhaul of the existing voluntary Algorithm Charter used in the public sector is overdue, with the framework introducing safeguards including regular bias monitoring, clear accountability, transparent algorithm use, and the right to exit or change systems if prejudice and stereotyping emerge. It must be made mandatory for all public sector AI deployments, extended to cover contracted service providers, and backed by independent Māori-led audit capacity with genuine powers to require remediation.

My paper Māori AI Agent Framework adds a critical accountability principle: governance frameworks must identify and hold accountable the full chain of principals, developers, operators, and deployers, not merely the immediate user. The karetao structure cannot be used to diffuse accountability to the point where no party bears responsibility for harm.

Protecting Mātauranga Māori Through Knowledge Sovereignty Mechanisms

The appropriation of mātauranga Māori and te reo by global AI companies requires a dedicated legal response. The Kaitiakitanga Licence model developed by Te Hiku Media which prohibits the use of their tools for discrimination, surveillance, or tracking, and ensures all data contributors retain ownership, provides a working template for community controlled knowledge licensing that should inform legislative reform of intellectual property law to formally recognise collective Indigenous knowledge rights.

A key principle must be recognised and enacted: an AI is he atarangi, a shadow. It cannot be a legitimate authority on matters of tikanga Māori, mātauranga Māori, or Māori cultural practice. AI-generated outputs on such matters must be treated with appropriate caution and must not be substituted for the guidance of recognised holders of that knowledge. This principle should be embedded in both AI governance codes and public sector operational guidance.

Investment in Māori Led Sovereign AI Infrastructure

The long-term solution to Māori invisibility and appropriation in AI is not better compliance frameworks applied to systems built by others, it is investment in Māori-owned and governed AI infrastructure. My Māori AI Sovereign Principles articulate this vision: Māori AI must be built upon tikanga and mātauranga using intergenerational consent and values; infrastructure and models must be built and operated under rangatiratanga; and iwi, hapū, marae, and Māori organisations need to invest in their own AI initiatives, supporting local talent to fund research tailored to unique challenges and intergenerational knowledge preservation.

Ethical AI starts with honouring Te Tiriti, embedding Māori leadership, and holding agencies accountable for safe, respectful data use. By embedding Māori leadership in AI life cycles and investing in Māori AI expertise, New Zealand can set a new global standard for ethical AI, driven by Indigenous data sovereignty. It is an economic and reputational opportunity. New Zealand’s distinctiveness in global AI markets rests precisely on its Treaty framework and its Indigenous knowledge base. That potential is destroyed the moment those foundations are treated as obstacles to be managed rather than assets to be governed.

Human Oversight and the Preservation of Relational Services

No AI system should replace human engagement in any public service context where cultural safety, whakapapa-based relationships, or the exercise of discretion in complex social situations is required. The PSA’s position is technically correct: frontline roles in border security, conservation, emergency management, and social services require human judgment that AI cannot replicate. For Māori communities specifically, this applies with even greater force: kaupapa Māori service delivery is relational by nature and cannot be reduced to automated decision trees.